The path of least resistance leads to the English language

Gareth Llewelyn Evans

The new Plaid Cymru government has committed additional resources to the NHS, childcare, free school meals and announced a National Productivity Goal spearheaded by Adam Price, whose responsibilities as Cabinet Minister for Enterprise, Connectivity and Energy include AI.

Price has spoken of AI’s potential to boost productivity – but has thus far failed to expand on the previous government’s work, leaving Welsh AI development and the use of the Welsh language within it at risk of being left behind.

The AI Plan for Wales, published last November by the outgoing Labour Government, rests on four pillars: economic growth, educating Wales, equitable delivery, and excellence and trust. The first two are where the investments sit. The third, the promise that AI in Wales will be fair and bilingual, and will leave no one behind. This is where the plan falls short.

AI is trained on data, whether that be text, audio, or images, with OpenAI’s GPT-3 trained on 570GB of plain text. English-Language Training data is plentiful, with ample audio recordings of conversations and an abundance of articles; however, the same cannot be said for the Welsh language. In total, there are around 249 hours of conversational Welsh audio training data, a far cry from the 1,217 hours of real conversational data used to train NVIDIA PersonaPlex. I have tried to train a Welsh speech-to-speech model myself and the data simply does not exist at a usable scale.

There are models that can hold a near-fluent conversation in Welsh. UK-LLM, built by University College London, Bangor University and NVIDIA, and trained on Isambard-AI — the UK’s most powerful supercomputer, is the best-performing major Large Language Model built to reason in Welsh alongside Scottish Gaelic, Irish, and Cornish. In north Wales, Bangor’s TechIaith has developed ‘Macsen’, a Welsh voice assistant that can tell you the news, the weather, and even play music on Spotify. Sounds like progress.

Yet both have drawbacks. Macsen, for example, takes a while to respond and follows a turn-based speech-to-speech system; compared to OpenAI Realtime, it is clunky and awkward to converse with. UK-LLM’s Welsh capabilities are developed via thirty million translated entries rather than original Welsh entries, which raises the question of what idioms, nuances and dialect forms are lost in translation. Would you pay £2,000 for hardware to run UK-LLM, or tolerate a clunky Macsen when you could get slightly broken Welsh from ChatGPT instead?

Meaningful adoption of Welsh-language AI requires developers to have access to public APIs (a technical interface which lets developers build on top of the technology). Rather than pay per message as is established with OpenAI or Anthropic, those who wish to use Welsh-language AI must set up their own infrastructure, potentially hindering public sector adoption – a direct contradiction of what the AI Plan for Wales aims for.

Commitment

Wales was the first country in the world to legislate for the well-being of future generations in 2015, with the AI plan citing the 2015 Act; every decision on AI must weigh its long-term impact. The same government introduced the ‘Cymraeg 2050’ goal, aiming to reach a million Welsh speakers by 2050. Yet the AI plan’s third pillar, equitable delivery, lacks the commitment to deliver on either.

If the AI tools people use on a day-to-day basis work fluently in English and remain clunky through the medium of Welsh, the path of least resistance pulls users to interact with the English-Language models. This dynamic undermines the intention of the ‘Cymraeg 2050’ initiative by the government’s own framework.

A factory cannot be built before the road is paved. Under this philosophy, it is unreasonable to expect public Welsh AI models to appear unless there is a concerted effort by Welsh institutions and the government to commission and publish open-source Welsh datasets – text, audio and translation data that universities, startups and public bodies could build on.

Training data

This does not require new investments on the scale of data centres. The Welsh government could secure compute time on the Isambard supercomputer to generate synthetic Welsh training data. It could fund public APIs for the models that already exist so that developers around Wales can integrate Welsh language AI into their applications as easily as English-language AI. The outgoing Labour Government explicitly left the new government to set its own priorities for ‘Cymraeg 2050’; in which direction will Plaid take it?

Silence, at this point in the new government, is a choice. As Cabinet Minister for Enterprise, Connectivity and Energy, Adam Price holds the responsibility; the fixes are cheap in comparison to what the government has already announced. Every day this continues, the path of least resistance carries another conversation into English.

---

Get more trusted Welsh news

Choose Nation.Cymru as a preferred source in Google News to see more of our journalism.